Before Buying AI Services for Your Company, Here Are the Questions to Ask the Vendor

Offers for AI services are everywhere now: chatbots, "AI agents", automation, internal assistants. As a business owner, it's hard to tell which ones are mature and which are just repackaged, especially when the demo always goes smoothly.

I build systems like this for clients myself, so treat this list as applying to me too. Good questions don't require you to understand the technical side. What you're looking for is an answer that's specific, in writing, and honest about the limits.

Why ask in this much detail

Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, because of escalating costs, unclear business value, or inadequate risk controls. Gartner also points to "agent washing", where vendors repackage chatbots, AI assistants, or RPA as agents, and estimates that only about 130 of the thousands of agentic AI vendors are real.

Regulators in the United States have also started acting on exaggerated AI claims. On March 18, 2024, the SEC fined two investment advisers, Delphia (US$225,000) and Global Predictions (US$175,000), for false claims about their use of AI. In September 2024, the FTC launched Operation AI Comply, and its chair at the time stressed that there is no AI exemption from the laws on the books.

Both examples are from the US, but the pattern is relevant: an "uses AI" claim is easier to write than to prove. That's why the questions need to be specific.

Eight questions for the vendor

Two extra questions I find useful: what can't this system do? and can I talk to a client who already uses it? An honest vendor usually answers the first one easily.

Three questions that often get skipped

Models get retired, so the system needs maintenance

AI models have a lifespan. On the Gemini API deprecations page (updated October 1, 2026), gemini-2.0-flash was released February 5, 2025 and is listed to shut down no earlier than June 1, 2026. The image model gemini-2.5-flash-image was released October 2, 2025 and is listed to shut down no earlier than October 2, 2026, exactly a year later. Google describes the dates in its table as the earliest possible dates, and promises to communicate the exact shutdown date with advance notice.

Anthropic promises at least 60 days' notice before retiring a publicly released model. For example, Claude Sonnet 4.5 was announced on September 30, 2026 and is scheduled to retire on November 30, 2026. The same page also notes that the temperature, top_p, and top_k parameters return an error if changed from their defaults on Claude Opus 4.7 and later. So switching models sometimes means changing code too.

That means an AI system isn't "built once and done". Someone has to track retirement schedules, retest the results on the new model, and bear the cost. That's why questions 2 and 8 are linked. My suggestion is to ask for a set of test cases (say, 20 real examples from your process) that can be rerun every time the model changes. This maintenance cost is also what people often miss when calculating ROI, as I cover in How to Calculate Automation ROI Before You Start the Project.

Is your data going through a free or a paid plan

Under the Gemini API terms, unpaid services, including the free quota and Google AI Studio, may use your content and the responses to develop Google products, and human reviewers may read and annotate them (the data is first disconnected from the account, API key, and Cloud project). For paid services, Google states it doesn't use prompts and responses to improve its products, and only logs them briefly to detect violations of its prohibited use policy.

That's why "which plan?" is a serious question. A vendor who runs a client's data through the free quota is making a decision on your behalf. For a comparison of personal and business plans at OpenAI, and what to check before pasting data, see Want to Summarize Your Company's Financial Data in ChatGPT? Is It Dangerous?.

"100% secure" is a warning sign

The OWASP Top 10 for LLM Applications 2025 puts Prompt Injection first (LLM01) and Excessive Agency, meaning a system given too much authority, sixth (LLM06). A serious vendor will bring up both without being asked. Even OpenAI, in its Dots safety FAQ, writes that its protections help reduce the risk of malicious instructions but don't eliminate it. If a company that size says the risk isn't zero, a vendor promising 100% secure deserves suspicion. I cover Dots itself in ChatGPT Now Has Dots, and What's Interesting Isn't How Smart It Is but Who Holds the Access.

On ownership, ask for a clear handover

Prompts, workflows, and the test set are assets, just like code. Make sure the accounts and API keys are in your company's name, and ask for a handover list: access, documentation, prompts, and test cases. The basic rules are similar to what I cover in You Don't Need to Understand Code, But You Need to Know This Before Paying a Developer and A Simple SOP to Keep Company AI a Team Asset, Not Personal Property.

How to use this list

1. Send the questions in writing to two or three vendors, and ask for written answers too.
2. Compare how honest they are about limits, not just the answers. A vendor who admits what the system can't do is usually more trustworthy.
3. Ask for a small trial with 20 real examples from your process (this is my suggestion), with pass criteria agreed first.
4. Put the important answers in the contract: ownership, data, maintenance, and costs.

A good vendor is usually glad to be asked this, because it shows you're serious. If you're weighing an AI offer and want a second opinion, tell me about your needs first.